Asplund spaces C_k(X) beyond Banach spaces

Put

Clearly, M is a bounded set. Using Proposition 1.1, we find a compact set K\subset X and a \gamma>0 such that

where

recall that this V^\gamma_K is a convex open neighbourhood of 0\in C_k(X). (This V^\gamma_K will witness for the Schwartz–Fréchet differentiability of \varphi at f).

As \varphi is Yamamuro–Fréchet differentiable at f, we have that

Hence

Now, consider any h\in V^\gamma_K. Then h[K]\subset (-\gamma,\gamma). Sine X is a Tychonoff space and K\subset X is compact, Proposition 1.5 yields a continuous extension \check h: X\to (-\gamma,\gamma) such that \check h|_K=h|_K. By the Claim below, we get that \varphi(f+th)=\varphi(f+t\check h), and therefore,

We also know that \lim_{t\to0}\frac1t(\varphi(f+th)-\varphi(f))=\varphi'(f)h for every h\in C_k(X). Thus, using the convexity of \varphi, we can conclude that

We proved that \varphi is Schwartz–Fréchet differentiable at the point f, where the neighborhood V^\gamma_K is a witness for that.

It remains to formulate and prove the announced claim/observation, going back to B. Sharp [s, p. 208, 209]:

Claim. If f,f'\in D and f|_K=f'|_K, then \varphi(f)=\varphi(f').

Proof of Claim. For all \lambda>1 we have f+\lambda(f'-f)\in D and

hence, by the convexity of \varphi, we have

Hence \varphi(f') \le \varphi(f). And, by symmetry, \varphi(f) \le \varphi(f').

We first realize that E admits an equivalent Fréchet smooth (off the origin) norm. Indeed, let \{x_1, x_2,\ldots\} be a countable dense subset of the unit ball S_E of E and \{\xi_1, \xi_2, \ldots\} be a countable dense subset of the unit ball S_{E^*} of the dual E^*. Then the assignment sending x^*\in E^* to

is an equivalent weak^* lower semi-continuous norm on E^*. For the reader’s convenience we indicate why |\cdot| is subadditive, i.e., the inequality |x^*+y^*| \le |x^*|+|y^*| holds. It is not easy to find in the literature how to prove this. A possible argument may profit from the subadditivity of the famous (Euclidean) \ell_2-norm.

A ‘small’ Kadec–Klee like effort, see [d, p. 113-117], reveals that the norm |\cdot| is locally uniformly rotund, i.e. if x^*, x^*_1,\ x^*_2,\ldots are elements of E^* such that 2|x^*|^2+2|x_i^*|^2-|x^*+x^*_i|^2\to0 when i\to\infty, then |x^*_i-x^*|\to 0 when i\to\infty. Assume this is not so. Find then an \varepsilon>0 and an infinite set N\subset {\mathbb N} such that \|x^*_i-x^*\| > 3\varepsilon for every i\in N.

Find m\in N so big that dist(x^*,{\rm sp}\{\xi_1,\ldots,\xi_m\}\big)<\varepsilon. A simple convexity argument provides an infinite set N_1\subset N such that dist(x^*_i,{\rm sp}\{\xi_1,\ldots,\xi_m\}\big)<\varepsilon for every i\in N_1. For every such i find y^*_i\in {\rm sp}\{\xi_1,\ldots,\xi_m\} such that \|x^*_i-y^*_i\|<\varepsilon, and so,

The sequence (y^*_i)_{i\in N_1} is bounded because \|x^*_i\|\to\|x^*\| as i\to\infty, by convexity. Further, since it lies in a finite-dimensional space, there are y^*\in {\rm sp}\{\xi_1,\ldots,\xi_m\} and an infinite set N_2\subset N_1 such that \|y^*_i-y^*\|\to0 as N_2\ni i\to\infty. Hence \|y^*-x^*\|\ge 2\varepsilon. Find n\in{\mathbb N} so that \|\langle y^*-x^*,x_n\rangle\| >\varepsilon.

Another simple convexity argument yields that \langle x^*_i,x_{n}\rangle \to \langle x^*,x_{n}\rangle as i\to\infty. Hence we have that

for all i\in N_2 big enough, a contradiction. We proved that our norm |\cdot| on E^* is LUR.

Now, having proved that our norm |\cdot| on E^* is LUR, the Šmulyan test guarantees that the predual norm |\cdot| on E, defined by

is Fréchet differentiable at every nonzero point of E. For more details, see [dgz, p. 43].

Next, let \overline x\in Df and \varepsilon>0 be given. We will find a v\in E such that |v-\overline x|<\varepsilon and \partial_F f(v) is nonempty. From the lower semi-continuity of f find \varepsilon'\in(0,\varepsilon) so small that B(\overline x,\varepsilon')\subset D and that f is bounded below on the open ball B(\overline x,\varepsilon'). Define

Then \varphi:E\longrightarrow[0,\infty] is easily seen to be proper and lower semi-continuous. Now, the Borwein–Preiss variational principle [p, Theorem 4.20] provides a Fréchet differentiable function \theta: D\longrightarrow[0,\infty) such that the sum f+\varphi+\theta attains infimum at some v\in D; clearly, v\in B(\overline x,\varepsilon'). We thus have

and so

We proved that -\varphi'(v)-\theta'(v)\in \partial_F f(v), and hence \partial_F f(v)\neq\emptyset.

First, we need a translation of Fréchet differentiability (of convex functions) to the terms of the space E: f is Fréchet differentiable at x\in E if and only if

this is easy to check. (Here and below, Q_+ means the set of all positive rational numbers). For any x\in E and any t>0, if S(x,t)<\infty, we find a vector u(x,t)\in B_E such that

(This u(x,t) is almost “the worst possible” regarding the Fréchet differentiability of f at x).

Let C_0 be a countable dense subset of Z. We construct countable sets C_0\subset C_1\subset C_2\subset\cdots\subset E as follows. Let m\in\mathbb{N} be given and assume that C_{m-1} was already found. Find a countable set C_m in E such that it is stable under making all finite linear combinations with rational coefficients, and that it contains C_{m-1} as well as the set \big\{u(x,t):\ x\in C_{m-1},\ t\in Q_{+}\}; clearly, C_m is again countable. Do so for every m\in\mathbb{N}, and put finally Y:=\overline{C_1\cup C_2\cup\cdots}\,. Clearly, Y\supset Z.

We claim that this Y has the desired property. So, assume that f|_Y is Fréchet differentiable at some x\in Y. We show that the whole f is Fréchet differentiable at x as well. (If x\in C_1\cup C_2\cup\cdots, then it is rather easy to proceed. So, we have to find an argument working also for x\in Y\setminus C_1\cup C_2\cup\cdots). Let L denote a Lipschitz constant of f in a vicinity of x; see [p, Proposition 1.6]. Pick any t\in Q_+ small enough (so that we can profit below from the L-Lipschitz property of f around x). Find then c\in\bigcup_{m=1}^\infty C_m such that \|c-x\|<t^2. We can now subsequently estimate for all sufficiently small t\in Q_{+} (so that we can use the Lipschitz property of f)

since f|_Y is Fréchet differentiable at x. Therefore, the “whole” f is Fréchet differentiable at x.

(1)\Longrightarrow(2) Let Y\subset C(X) be any separable subspace. Choose a countable and linearly independent set N\subset B_Y such that its linear span is dense in Y.

Consider a correspondence

Thus \psi: X\longrightarrow [-1,1]^{N}, and this is a continuous mapping. Denote L:=\psi(X); this is a metrizable compact space. And, as X is scattered, L is scattered as well, by Theorem 1.3. Moreover, L is countable, by Theorem 1.4.

Next, let us enumerate L as \{z_{n}:\ n\in N\}. It is known that every linear continuous functional x^*\in C(L)^{*} has the form

and \|x^*\|=\sum_{n\in N} |a_n| < \infty, where (a_n)_{n\in N} is a suitable element of \ell_1(L), see [Pelczynski-Semadeni, Main Theorem (11)] or [Semadeni, Corollary 19.7.7]. Consequently, C(L)^{*} is isometric to \ell_1(L), and this implies that C(L)^* is separable.

We will show that the subspace Y\subset C(X) is isometric to a subspace of C(L). For n\in N and x\in X we put

The \tilde n is a well defined function on L. Indeed, if \psi(x)=\varphi(h) for some x, h\in X, then \tilde n(\psi(x))= n(x)=n(h)=\tilde n(\psi(h)).

Fix any n\in N. The function \tilde n belongs to C(L). Indeed, consider a net (\psi(x_\alpha)) in L converging to some \psi(x)\in L. This means that m(x_\alpha)\to m(x) for every m\in N. Thus, in particular, n(x_\alpha)\to n(h). Therefore, \tilde n(\psi(x_\alpha))=n(x_\alpha)\longrightarrow n(x)= \tilde n(\psi(x)).

Now, for any element (a_n)_{n\in N} \in c_{00}(N) we have

(In both C(L) and C(X), we considered the maximum norm \|\cdot\|). Hence the mapping

is well defined, and it is a linear isometry. Also, we know that the linear span sp(N) is dense in Y. Thus, we can uniquely extend this mapping to the linear isometry defined on the whole Y into C(L), say T: Y\hookrightarrow C(L).

Finally, the operator T it injective, a simple consequence of the Hahn–Banach theorem reveals that the adjoint operator T^*: C(L)^*\to Y^* is surjective. And recalling that C(L)^* was separable, Y^* must be separable as well by Hahn-Banach Theorem.

(2)\Longrightarrow(3) Is clear since the dual of \ell_1 is isometric to the non-separable space \ell_\infty\,.

(3)\Longrightarrow(1) Assume on the contrary that X is not scattered. By Theorem 1.3, there is a continuous surjection \rho: X\twoheadrightarrow [0,1]. Then the adjoint mapping \rho^{*}: C[0,1]\hookrightarrow C(X) is a linear isometry into. But the Banach space C[0,1] is universal for the class of all separable Banach spaces by the Banach–Mazur theorem, see [Pelczynski-Bessaga, Proposition 1.5]. Therefore, C(X) contains an isometric copy of \ell_1. Obtained contradiction completes the proof.

The necessity will be proved in a more general Theorem 3.2. Just replace there X by L in the proof of implication ‘(1)\Longrightarrow(2)’.

Sufficiency. Assume that X is scattered. Take any open convex set D\subset C(X) and any convex continuous function \varphi: D\to{\mathbb R}. We have to show that \varphi is Fréchet differentiable at the points of a dense subset of D. So fix any open set U\subset D. Pick some separable subspace Z\subset C(X) such that U\cap Z is a nonempty set. For this Z, we find a separable over(sub)space Z\subset Y\subset C(X) as it is done in Proposition 2.3.

As X is scattered, Theorem 2.4 guarantees that the dual Y^* is separable. Hence, Y is an Asplund space by Proposition 2.1. Thus, there is a point f in the set U\cap Y (open in Y) where the restriction \varphi_{|Y} is Fréchet differentiable. Now, the conclusion of Proposition 2.3 says that the whole function \varphi is Fréchet differentiable at f. We proved that the set of points where the function \varphi is Fréchet differentiable is dense in D. Finally, Proposition 1.10 completes the proof.

Let S:Y\hookrightarrow C_k(X) be the promised isomorphism into. Then there must exist some \varepsilon>0, \alpha>0, and a compact subset K\subset X such that

We show that the (Banach) space Y is isomorphic to a subspace of the (Banach) space C(K). Define \tilde S:Y\to C(K) by

\tilde S is (linear and) injective. Indeed, take any y\in Y such that \tilde Sy=0. This means that (Sy)|_K\equiv0. Then, for every n\in {\mathbb N} we have

by (Proof), and so Sy\in S[\frac1{n} \stackrel \circ B_Y], and hence y\in \frac 1{n} \stackrel \circ B_Y, as S is injective. Therefore, y=0.

Now, fix any y\in \stackrel \circ B_Y. By (Proof) Sy \in V^{\varepsilon/\alpha}_K, i.e. Sy[K] \subset (-\varepsilon/\alpha,\varepsilon/\alpha), and hence \|\tilde Sy\| <\varepsilon/\alpha. We thus proved that

It remains to prove that \varepsilon \stackrel \circ B_{C(K)}\cap \tilde S[Y] \subset \tilde S[\stackrel \circ B_Y]. Fix any y\in Y such that \tilde Sy \in \varepsilon \stackrel \circ B_{C(K)}; thus \|\tilde Sy\| <\varepsilon. Hence Sy\in V^\varepsilon_K \cap S[Y]. By (Proof), there is a y'\in \stackrel\circ B_Y so that Sy=Sy'. But S is injective. Hence y=y' and so y\in \stackrel \circ B_Y. We proved that \varepsilon \stackrel \circ B_{C(K)} \cap \tilde S[Y] \subset \tilde S[\stackrel \circ B_Y], and therefore, \tilde S is an isomorphism of Y onto \tilde S[Y] \ \ (\subset C(K)). Summarizing, we obtained that

that is, analytically (using the injectivity of \tilde S),

(1)\Longrightarrow(2) Assume that C_k(X) is an Asplund space and that some compact subset set L\subset X is not scattered. Find then a nonempty closed perfect set P\subset L, by Lemma 1.2. Consider the function \varphi: C_k(X)\to{\mathbb R} defined by

Clearly, \varphi is convex. It is also continuous on C_k(X). Indeed, pick any f\in C_k(x) and any \varepsilon>0. Then

for every h\in V_P^\varepsilon we have

and, similarly,

thus |\varphi(f+h)-\varphi(f)| <\varepsilon.

Put

this is obviously a bounded set. As C_k(X) is Asplund and \varphi is continuous convex, \varphi is Fréchet differentiable at some f\in C_k(X). Thus, in particular, there is a linear continuous \xi: C_k(X)\rightarrow{\mathbb R} such that

and hence (getting rid of the almost redundant \xi)

Thus, in particular, (for \varepsilon:=\frac12) there is a \delta>0 such that

for every h\in M.

Now, pick a p\in P so that f(p)=\varphi(f). As P is perfect, there is a q\in P\setminus \{p\} such that f(q)> \varphi(f)-\frac\delta 2.

We recall that X is a Tychonoff space and L a compact set in it. Hence, there is a continuous function h: X\to [0,1] such that h(p)=0 and h(q)=1. Note that then \varphi(h)=1 and h\in M.

Thus

(Note that we used a Šmulyan like argument, see [dgz]) and [y2, p. 342–344]; a contradiction. We proved that L must be scattered.

(2)\Longrightarrow(1) Here we freely imitate arguments from [s] and [es]. Let D\subset C_k(X) be any (nonempty) open convex set and consider any convex function \varphi: D\to {\mathbb R}, everywhere continuous in the k-topology of the space C_k(X). We want to show that the set of points where \varphi is Fréchet differentiable contains a dense G_\delta subset of D.

Pick some u\in D. The continuity of \varphi at u yields an \varepsilon>0 and a compact subset L\subset X such that u+V^\varepsilon_L \subset D and that \varphi is bounded above on the set u+V^{\varepsilon}_L. A simple picture, with [p, p. 4, 5], then reveals that for every v\in D there is a \gamma>0 such that v+ V^{\gamma}_L \subset D and that \varphi is bounded above on v+ V^{\gamma}_L.

Let C(L) be the Banach space of all continuous functions on L, with ‘maximum’ norm. Its closed unit ball (around the origin) is denoted by B_{C(L)}. Consider the ’restriction’ mapping T: C_k(X)\to C(L) defined by

Clearly, T is linear. It is also continuous since, for every \gamma>0, we have T[V^\gamma_L]\subset \gamma\!\!\stackrel \circ B_{C(L)}. (We have here even equality by Proposition 1.5). Since T is continuous, T^{-1}(\Omega) is open or G_\delta whenever \Omega\subset C(L) is, respectively, open or G_\delta. Let us further check that T[D] is an open subset of C(L). So, fix any f\in D. Find a \gamma>0 so that f+ V^\gamma_L\subset D. Then

the equality here comes from Proposition 1.5, as X is Tychonoff and L is compact. We verified that T[D] is open.

We further observe that, given any dense subset \Omega in T[D], then T^{-1}[\Omega] is dense in D. Indeed, pick any v\in D, any \beta>0, and any compact set L' in X. We want to show that (v+V^\beta_{L'})\cap T^{-1} [\Omega]\neq\emptyset. Take \gamma\in(0,\beta) so small that v+V^{\gamma}_L\subset D.

As \Omega is dense in T[D], we have that

hence, there is an h\in \gamma\!\! \stackrel \circ B_{C(L)} so that Tv+h\in\Omega. By Proposition 1.5, we find a continuous function \check h:X\to{\mathbb R} such that \check h[X] \subset (-\gamma,\gamma) and \check h|_L=h. Then v+\check h \in v + V^{\gamma}_{L'} \subset v + V^{\beta}_{L'} and T(v+\check h)=Tv + h\ (\in\Omega). Thus

We proved that T^{-1}[\Omega] is dense in D; note that T^{-1}[\Omega]\cap D is dense in D, too.

Now, we define the function \psi: T[D]\to{\mathbb R} as

where \check g is any element of D such that \check g|_L=g. This is a correct definition because, if f,f'\in D and f|_L=f'|_L=g, then by the Claim from the proof of Proposition 1.6, \varphi(f)=\varphi(f').

The function \psi is convex. Indeed, take any g_1,g_2\in {C(L)} and any \alpha\in (0,1)

Find some \check{g_1}, \check{g_2}\in D, with \check{g_1}|_L=g_1 and \check{g_2}|_L=g_2. Then

and hence

Let us check that our \psi is continuous on T[D]. Fix any f\in D. We know that there is an \varepsilon>0 so that f+V^\varepsilon_L lies in D and that \varphi[f+ V^\varepsilon_L] is bounded above. Thus the set

is also bounded above (we needed Proposition 1.5). And having this, then [p, p. 4, 5] (valid for Banach spaces) guarantees the continuity of \psi.

Now, since L is a scattered compact set by the assumption, Theorem 2.5 says that the Banach space C(L) is Asplund. Hence our \psi is Fréchet differentiable at the points of a dense G_\delta subset, say \Omega, of T[D].

Fix any f in T^{-1}(\Omega)\cap D (which is already known to be a dense G_\delta subset of D). Find an \varepsilon>0 so small that f+V^\varepsilon_L\subset D (we already proved that such an \varepsilon exists). As Tf\in\Omega, the function \psi is Fréchet differentiable at Tf, with derivative, \xi: C(L)\to{\mathbb R}, say. Having this we get

Here we again used Proposition 1.5. Therefore, our \varphi is Fréchet differentiable at f, with derivative

Actually we should in addition check that this \varphi'(f) is linear and continuous. Indeed, we have that \xi[V^\varepsilon_L] \subset (-\varepsilon,\varepsilon).

(3)\Longrightarrow(2) Assume that X contains a compact subset K which is not scattered. By Theorem 1.3, there exists a continuous surjection \rho: K\twoheadrightarrow [0,1].

Using Proposition 1.5, \rho can be extended to a continuous surjection, say \check \rho: X\twoheadrightarrow [0,1].

The natural adjoint mapping to \check\rho, say T: C[0,1]\hookrightarrow C_k(X), is then an embedding (i.e. an isomorphism onto its range). Indeed, a bit of routine calculation reveals that

(We note that the above equality says that our T is a ‘bound-covering mapping’, see [es]). Since C[0,1] is universal in the class of all separable Banach spaces, by [Pelczynski-Bessaga, Proposition 1.5], the Banach space \ell_{1} (isometrically) embeds into C[0,1]. Hence \ell_1 embeds into C_k(X), which is a negation of (3).

(2)\Longrightarrow(3) Let (2) hold. Assume that C_{k}(X) contains an isomorphic copy of \ell_{1}. By Lemma 3.1 there exists a compact (scattered) subset K\subset X such that \ell_1 is isomorphic to a subspace of the Banach space C(K). But his contradicts Theorem 2.4. Hence C_k(X) does not contain any isomorphic copy of \ell_1.

(4)\Longrightarrow(3) It is true since the dual of \ell_1 is isometric to \ell_\infty, which is non-separable.

(2)\Longrightarrow(4) It follows from Theorem 2.4 and Lemma 3.1.

(1)\Longleftrightarrow(2) This is Lemma 4.3.

(1)\Longrightarrow(3) Trivially follows from Theorem 4.4(a).

(3)\Longrightarrow(2) Assume that X is a \Delta_1-space. Then every compact subset of X is a \Delta_1-space. We apply Theorem 4.4(b).

Assume on the contrary that each compact subset in X is scattered but X is not scattered. Since the space X is pseudocompact, we can apply [Leiderman-Tka, Proposition 5.5] to get in X a closed subset F for which there exists a continuous surjection h: F\twoheadrightarrow [0,1]. Let Q be a dense countable subset of [0,1]. Choose a countable subset C\subset F such that h(C)=Q. Let A:=\overline{C} be the closure in F. Note that F is also \omega-bounded. Therefore, A is compact and obviously h(A)=[0,1]. By assumption, the compact set A is scattered, this is a contradiction, in view of Theorem 1.3. Hence X is scattered. The converse is clear.

Every compact subset of Q is countable, hence it is scattered. By Theorem 3.2 we deduce that C_k(Q) is an Asplund space. Since Q is a \mu-space, the space C_k(Q) is barrelled by the Nachbin–Shirota theorem, see [KKPS, Propostion 2.4.16], i.e., every closed absolutely convex absorbing subset of C_k(Q) is a neighbourhood of zero.

On the other hand, the space C_k(Q) is not Baire-like, i.e., there exists an increasing sequence of closed convex absorbing subsets A_1, A_2,\ldots covering the space C_k(Q) such that no A_n is a neighbourhood of the origin in C_k(Q). Indeed, assume on the contrary that C_k(Q) is a Baire-like space. Then we follow the argument contained in the proof of [KKPS, Proposition 2.4.20], or [mccoy, Corollary 5.3.4]: For n\in{\mathbb N} we put

Clearly, the U_n’s form a basis of neighbourhoods of 0 in Q and the (closed convex and absorbing) sets H_n’s cover all of C_k(Q). Also, clearly, U_1 \supset U_2\supset \cdots and H_1\subset H_2\subset \cdots.

Since we assumed that C_{k}(Q) is Baire-like, there exist a compact set K\subset Q, \varepsilon>0, and n\in{\mathbb N} such that

Note that U_{n}\subset K. Indeed, if there is a z\in U_{n}\setminus K, then putting

we have that f\in C_k(Q) and that f(z)=2n. Also f_{|K}\equiv 0, and thus f\in U_n, which is impossible. Therefore, we have that U_{n}\subset K. This leads to a contradiction because K is a countable compact subset of {\mathbb R} and the closure of U_{n} in {\mathbb R} is an uncountable closed interval. We have proved that the space is C_k(Q) is not Baire-like. Finally, since C_k(Q) is barrelled but not Baire-like, we apply the Saxon’s result (see [KKPS, Corollary 2.4.4]), and conclude that lcs C_k(Q) contains isomorphically a (closed) copy of \varphi.